Miles is an open-source framework from RadixArk built to address the distributed-systems challenges of large-scale reinforcement learning post-training for LLMs. It composes SGLang for rollout generation, NVIDIA Megatron-LM for training, Ray for orchestration, and PyTorch for core numerical computation. The framework provides production features including fast weight synchronization over NCCL/RDMA, MoE-aware routing alignment, unified low-precision support, fault tolerance, and pluggable components for custom rollout, reward, and loss functions.
